Kon: A Minimalist Coding Agent

A developer has released Kon, a coding agent designed for local systems. The main goal is to provide an essential tool that is easy to understand, customize, and extend.

Kon stands out for its small size: the system prompt requires approximately 215 tokens, and the tool definitions require around 600 tokens, keeping the total footprint below 1000 tokens before the conversation context. The repository comprises about 108 files, which, according to the developer, makes it easy to understand in a weekend.

Implementation Details

The agent was tested with a glm-4.7-flash-q4 model on a machine with an i7-14700F processor, 64 GB of RAM, and an RTX 3090 GPU with 24 GB of VRAM. Kon draws inspiration from the pi-coding-agent project.

For those evaluating on-premise deployments, there are trade-offs to consider carefully. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.